Thoughts and experiments around AI and coding agents
14 entries
Claude Code's two multi-agent modes: a coordinator with disposable workers, and a swarm of teammates in terminal panes.
Once a day, Claude Code forks a background agent to consolidate its memories — the source calls it dreaming.
Claude Code speculatively runs your predicted next turn while you're idle, then commits it if you accept.
Claude Code's two hidden defenses — one against model distillation, one against leaking internal codenames.
How Claude Code guards a tool call: an auto-mode AI classifier, fail-closed gates, and 23 Bash security rules.
Five compression strategies, applied lightest to heaviest, keep Claude Code's context window from overflowing.
Claude Code fights long-session context drift with a hot-warm-cold, three-tier memory architecture.
Claude Code skips RAG entirely, letting the agent grep and a small metadata index handle retrieval.
Claude Code splits its system prompt at a sentinel boundary to keep the cacheable prefix stable.
Claude Code runs tools concurrently using the read/write-lock pattern straight out of databases.
How Claude Code's 40+ built-in tools are registered, deferred, and locked down with fail-closed defaults.
The autonomous agent at the heart of Claude Code is, underneath, a single while(true) loop.
An end-to-end tour of Claude Code's architecture, and why it's all fundamentals rather than secret sauce.
Learn advanced features, shortcuts, and workflows to get the most from Claude Code.
